Researchers introduce Controllable Neural Variational Agents (CNeVA), a framework that learns per-agent Gaussian behavior latents from discounted returns to enable steering of simulated agents along interpretable axes. The system conditions a rectified-flow trajectory generator and employs soft eligibility gates to preserve gradient signals where hard thresholds would fail.

  • CNeVA infers behavior latents via closed-form conjugate variational updates on the Waymo Open Motion Dataset.
  • Soft eligibility gates replace binary thresholds with exponential decay to maintain gradient flow for near-threshold agents.
  • The model achieves competitive realism while providing per-channel controllability that higher-ranked imitation models lack.
  • Steering metrics must be evaluated alongside physical-plausibility guardrails to avoid reward-hacking confounds.

This approach allows engineers to isolate variables and reproduce specific edge cases for testing autonomous systems without real-world risk.